I’m a Kaggle Grandmaster!

I’ve achieved 𝗞𝗮𝗴𝗴𝗹𝗲 𝗚𝗿𝗮𝗻𝗱𝗺𝗮𝘀𝘁𝗲𝗿 status! At present, there are only 464 of us worldwide. I’m also part of an even smaller group: 𝗞𝗮𝗴𝗴𝗹𝗲 𝗖𝗼𝗱𝗲 𝗚𝗿𝗮𝗻𝗱𝗺𝗮𝘀𝘁𝗲𝗿𝘀—just 84 at last count.

If you’re unfamiliar, Kaggle (http://kaggle.com) is a 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 where people share code and compete in challenges on topics ranging from diagnosing disease using sensor data to using generative AI to fix bugs on GitHub.

Code Grandmaster status is earned by sharing solutions to machine learning problems (as Python notebooks) that are voted “Gold” by expert-level Kaggle users.

Out of over 50,000 Kaggle users, 𝗺𝘆 𝗰𝗼𝗱𝗶𝗻𝗴 𝗰𝗼𝗻𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻𝘀 𝗰𝘂𝗿𝗿𝗲𝗻𝘁𝗹𝘆 𝗿𝗮𝗻𝗸 𝟱𝘁𝗵 𝗴𝗹𝗼𝗯𝗮𝗹𝗹𝘆.

👉 You can check out my Kaggle notebooks here: https://www.kaggle.com/richolson/code

Kaggle competitions often come with prize money—typically around $𝟭𝟱,𝟬𝟬𝟬 𝗳𝗼𝗿 𝗳𝗶𝗿𝘀𝘁 𝗽𝗹𝗮𝗰𝗲. But the best part? There are no barriers to entry. You don’t have to start from scratch—you can take someone else’s code, build on it, and improve it. This gets to how I became a Code Grandmaster.

When a new competition drops, I get to work on a 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻: a relatively simple approach that shows how the problem can be meaningfully solved. Every competition has a public leaderboard based on a scoring metric. A decent score means I’m on the right track.

Once I have a working solution, I go back and streamline it—boiling it down to the core ideas and removing anything unnecessary. I add notes to explain parts that aren’t immediately obvious, aiming to make the key concepts understandable to someone with only minimal machine learning experience.

Then I 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝗻𝗼𝘁𝗲𝗯𝗼𝗼𝗸 𝗽𝘂𝗯𝗹𝗶𝗰—for anyone to copy and improve on.

Some Examples:

🖐️ Using a 1D CNN to decipher gestures from smartwatch-style sensor data:
https://www.kaggle.com/code/richolson/cmi-2025-1d-cnn-imu-only-baseline

🧪 Predicting physical properties of polymers from chemical structure:
https://www.kaggle.com/code/richolson/smiles-rdkit-lgbm-ftw

🧠 Using small LLMs to recognize errors in human math:
https://www.kaggle.com/code/richolson/eedi-llm-benchmark

🖼️ Classifying skin lesions as cancerous or benign using an ImageNet:
https://www.kaggle.com/code/richolson/isic-2024-imagenet-train-oof-preds-public

🕶️ Making LLMs misbehave with adversarial prompt engineering:
https://www.kaggle.com/code/richolson/add-it-up

Sharing most of my work isn’t exactly conducive to winning competitions. People can easily tweak my public notebooks and outscore me with my own code.

So why did I author over 100 public notebooks if they hurt my chances at prize money?

Because I wasn’t (usually) trying to win—𝗜 𝘄𝗮𝘀 𝗱𝗼𝗶𝗻𝗴 𝗶𝘁 𝘁𝗼 𝗹𝗲𝗮𝗿𝗻.

“If you can’t explain it simply, you don’t understand it well enough.”
— 𝗗𝗼𝗹𝗹𝘆 𝗣𝗮𝗿𝘁𝗼𝗻
(Not really. But also 𝗻𝗼𝘁 𝗔𝗹𝗯𝗲𝗿𝘁 𝗘𝗶𝗻𝘀𝘁𝗲𝗶𝗻.)